WIP: harbor backend v2

[mathewjhan]

What

Why

How to Test

Checklist

• PR title follows [module] type: description format
• Appropriate labels added (e.g. enhancement, bug, breaking)
• ruff check . and ruff format --check . pass
• pyright osmosis_ai/ passes
• pytest passes (new tests added if applicable)
• Public API changes are documented
• No secrets or credentials included

---

Summary by cubic

Add HarborBackendV2 to run Harbor tasks with pure task images by installing a bundled workflow per trial or using native Harbor agents. Includes a small container I/O contract, diagnostics with secret redaction, artifact relocation, and an end‑to‑end benchmark.

• New Features

  • HarborBackendV2: runs bundled workflows or native agents (terminus-2, mini-swe-agent, oracle); supports template/dataset modes and per‑phase timing with failure diagnostics; redacted diagnostics exposed via extra_fields on samples.
  • Bundling (osmosis_ai.packaging): builds a wheel from a pyproject.toml harness, injects agent/grader console scripts, caches by content hash in the user cache dir via platformdirs, and exposes inspect_bundle/project_dir_for.
  • Container contract/runner: ContainerInput/ContainerResult plus in‑container runner.agent_main/runner.grader_main; extracts messages from trajectories, writes reward files for Harbor, and adds AgentWorkflowOutput.
  • Task materialization: HarborTask + TaskMode copy per rollout, write instruction.md, patch Dockerfiles to install SDK deps into an isolated venv (/opt/osmosis/venv) bootstrapped with uv (fallback to pip), and generate tests/test.sh to install a grader when needed.
  • Native agent wiring: passes model endpoint via env/kwargs per agent; bundled workflows run from the venv.
  • HarborBackend: content‑addressed image tags to reuse builds; new exports in osmosis_ai.rollout.backend.harbor.__init__.
  • Extras: unit tests for packaging and backend; benchmarks/container_lifecycle to measure cold/warm throughput.

• Migration

  • Use HarborBackendV2 with tasks_dir; choose a native agent by name or provide a workflow harness (code_dir or a prebuilt bundle).
  • Ensure the harness is a Python project with a pyproject.toml; the bundler creates <package>-agent/<package>-grade scripts automatically.
  • Bundles install into /opt/osmosis/venv; images only need Python. The agent installs uv if missing and then the bundle (falls back to pip). For dataset runs, set task_mode="dataset" and include harbor_task_id in request metadata.

^{Written for commit c52aa63. Summary will update on new commits.}

[codecov]

Codecov Report

❌ Patch coverage is 55.85938% with 113 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
osmosis_ai/rollout/container/runner.py	0.00%	89 Missing ⚠️
osmosis_ai/packaging.py	88.98%	6 Missing and 7 partials ⚠️
osmosis_ai/rollout/container/trajectories.py	30.76%	9 Missing ⚠️
osmosis_ai/rollout/container/files.py	94.44%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!